Depth completion, the technique of estimating a dense depth image from sparse depth measurements, has a variety of applications in robotics and autonomous driving. However, two main challenges depth completion faces are the irregularly spaced pattern in the sparse depth input, and the lack of dense pixel-level ground truth depth labels for training. Those challenges turn to be even more complicated when it comes to data from the underwater domain. In my work I concentrate in developing a training framework which learns a direct mapping from sparse depth estimations and color images input to dense depth prediction - in a self supervised manner.